Developing a General Framework for Human Autonomy Teaming...–radar altimeter –medical emergency –heart attack –laceration –engine loss –depressurization –damage –cabin

Developing a General Framework for

Human Autonomy Teaming

Joel Lachter

Summer L. Brandt

R. Jay Shively

April 18, 2017

1

Problems with Automation

• Brittle

– Automation often operates well for a range of situations but requires human

intervention to handle boundary conditions (Woods & Cook, 2006)

• Opaque

– Automation interfaces often do not facilitate understanding or tracking of the system

(Lyons, 2013)

• Miscalibrated Trust

– Disuse and misuse of automation have lead to real-world mishaps and tragedies (Lee

& See, 2004; Lyons & Stokes, 2012)

• Out–of-the-Loop Loss of Situation Awareness

– Trade-off: automation helps manual performance and workload but recovering from

automation failure is often worse (Endsley, 2016; Onnasch, Wickens, Li, Manzey,

2014)

2

Tenets of Human Autonomy Teaming (HAT)

Transparency

Communication of Rationale

Communication of Confidence

Shared Language

Shared Goals

Shared Plans

Agreed allocation of responsibility

Minimized Intent Inferencing

3

Bi-Directional

CommunicationPlays

Make the Automation into a Teammate

HAT Agent

4

Implementation

5

Tenets Human In The Loop

Simulations

Implementation

6

Tenets Human In The Loop

Simulations

Simulated Ground Station

7

ELP and ACFP

ELP – Emergency Landing Planner (2007-2012)

– Cockpit decision aid

– Route planning for (serious) emergencies

– control system failures

– physical damage

– fires

– Time & Safety were dominant considerations

ACFP – Autonomous Constrained Flight Planer (2013-2017)

– Ground station decision aid

– Diversion selection, route planning, route evaluation

– weather diversion

– medical emergencies

– less critical system failures

Research prototype software, Intelligent Systems Division, PI: D. Smith

Find the best landing sites and routes for the aircraft

ELP Objective

Icing

damage/failures

recovery

Runway

length/width/conditionPopulation

Facilities

En route

Weather

Distance

Wind

Altitude

Ceiling, Visibility

Approach

ELP Approach

Consider all runways within range (150 miles)

Construct “obstacles” for weather & terrain

Search for paths to each runway

Evaluate risk of each path

Present ordered list

< 10 seconds

ELP’s Risk Model

Enroute path

Distance/time

Weather

Approach path

Ceiling & Visibility

Approach minimums

Population density

Runway

Length

Width

Surface condition

Relative wind

Airport

Density altitude

Tower

Weather reporting

Emergency facilities

Pstable ≡ probability of success / nm in stable flight

Pwx ≡ probability of success / nm in light weather

Pleg ≡ (Pstable ∗ (Pwx )S )D

Proute ≡ ∏ Pleg

Icing

Pappr ≡ Pleg ∗ Pceil ∗ Pvis

Prnwy ≡ Plength ∗ Pwidth ∗ Psurf ∗ Pspeed ∗ Pxwind

1

0Reqd

length

Plength

Emergency Page on the CDU

Airport

Runway length

Distance to airport

Bearing to airport

Page #

Select Show Airport Info Page

Update

Runway

Principal Risks

Go to Previous/Next Page

Execute the selection

ELP Routes on the Navigation Display

ELP Experiment (2010)

Evaluation of ELP in ACFS– 3 physical damage scenarios

– 5 pilot teams

– 16 scenarios each

Results– Decision quality somewhat better in adverse weather

– Decision speed much better in adverse weather

– Damage Severity not a significant factor

Pilot feedback:“ ... your software program alleviates the uncertainty about finding a suitable

landing site and also reduces workload so the crew can concentrate on "flying" the aircraft.”

The Emergency Landing Planner Experiment

Nicolas Meuleau, Christian Neukom, Christian Plaunt, David Smith & Tristan Smith

ICAPS-11 Scheduling and Planning Applications Workshop (SPARK), pages 60-67, Freiburg, Germany, June 2011

ACFP differences

Multiple aircraft

Much wider geographic area

Additional optimization criteria

– medical facilities

– maintenance facilities

– passenger facilities

– connections

Constrained requests

– runway length

– distance

Route evaluation

– current route/destination

– proposed changes

RCO Ground station

Optimization

Situations:

– weather reroute

– weather diversion

– systems diversion

– anti-skid braking

– radar altimeter

– medical emergency

– heart attack

– laceration

– engine loss

– depressurization

– damage

– cabin fire

Safety Time Medical Conven. Maint.

Simulated Ground Station

17

Implementing HAT Tenets in the Ground Station

18

Implementing HAT Tenets in the Ground Station

19

Implementing HAT Tenets in the Ground Station

• Human-Directed: Operator calls “Plays” to determine who does what

20

A play encapsulates a plan for

achieving a goal.

It includes roles and responsibilities

what is the automation going to

do

what is the operator going to do

Implementing HAT Tenets in the Ground Station

• Transparency: Divert reasoning and

factor weights are displayed.

• Bi-Directional Communication:

Operators can change factor weights to

match their priorities. They can also

select alternate airports to be analyzed

• Shared Language/Communication:

Numeric output from ACFP was found

to be misleading by pilots. Display now

uses English categorical descriptions.

21

HAT Simulation: Tasks

• Participants, with the help of automation, monitored 30 aircraft

– Alerted pilots when

• Aircraft was off path or pilot failed to comply with clearances

• Significant weather events affect aircraft trajectory

• Pilot failed to act on EICAS alerts

– Rerouted aircraft when

• Weather impacted the route

• System failures or medical events force diversions

• Ran with HAT tools and without HAT tools

22

HAT Simulation: Results

• Participants preferred the HAT condition overall (rated 8.5 out of 9).

• HAT displays and automation preferred for keeping up with operationally

important issues (rated 8.67 out of 9)

• HAT displays and automation provided enough situational awareness to

complete the task (rated 8.67 out of 9)

• HAT displays and automation reduced the workload relative to no HAT (rated

8.33 out of 9)

23

HAT Simulation: Debrief

• Transparency

– “This [the recommendations table] is wonderful…. You would not find a dispatcher

who would just be comfortable with making a decision without knowing why.”

• Negotiation

– “The sliders was [sic] awesome, especially because you can customize the route…. I

am able to see what the difference was between my decision and [the computer’s

decision].”

• Human-Directed Plays/Shared Plans

– “Sometimes [without HAT] I even took my own decisions and forgot to look at the

[paper checklist] because I was very busy, but that didn’t happen when I had the

HAT.”

24

HAT Simulation: Summary

• Participants liked where we were headed with the HAT concept

– Increased Situation Awareness

– Reduced Workload

• Things we didn’t get quite right

– Annunciations: People liked them but thought there were to many

– Voice Control: Did not work well. Need a more complete grammar, better recognition

– Participants didn’t always understand what the goal of a play was

• Things we didn’t get to

– Airlines hate diverts. We need to put in support to help avoid them

– Plays need more structure (branching logic)

– Roles and responsibilities need to be more flexible

– Limited ability to suggest alternatives

25

Summer ’17

Generalization

26

Tenets Human In The Loop

Simulations

Generalization

27

Tenets Human In The Loop

Simulations

Thought

Experiments

HAT in Photography

28

HAT in Photography

29

HAT in Photography

30

HAT in Photography

31

HAT in Photography

32

HAT in Photography

33

HAT in Navigation

34

HAT in Navigation

35

HAT in Navigation

36

Lessons

• Seems applicable to a

wide variety of

automation

• Plays are a big part of the

picture

– Provide a method for

moving negotiation to

less time critical periods

– Provide a mechanism for

creating a shared

language

37

Tenets Human In The Loop

Simulations

Thought

Experiments

Design Patterns

• Looking at a variety of situations, we see common problems with common

solutions

– Bi-Directional Communication solves a problem of keeping the human in the loop with

potential problems in the current plan and reduces brittleness by opening up the

system to operator generated solutions

– Plays solve the problem allowing the system to adopt to different conditions without

having the system infer the operator’s intent

• In other domains, people have attempted to capture similar problem-solution

pairs using “design patterns”

– Architecture and Urban Planning (Alexander, et al., 1977)

• E.g., Raised Walkways solve the problem of making pedestrians feel comfortable

around cars

– Computer Programming (Gamma, et al., 1994)

• E.g., Observers solve the problem of maintaining keeping one object aware of

the state of another object

38

Design Patterns for HAT

• Working with the NATO working

group on Human Autonomy

Teaming (HFM-247) to develop

design patterns for HAT

• Original Conception was to

identify relationships between

different agents (after Axel

Schulte, Donath, & Lange,

2016)

39

Design Patterns for HAT

• Working with Gilles Coppin from the

NATO Working Group on a Bi-

Directional Communication pattern

• Modeled after Gamma et al

specifications:

– Intent: Support generation of input

from all relevant parties and its

integration into decisions

– Motivation: Reduce brittleness of the

system by consolidating information

and skills

– Applicability: May not be applicable in

urgent situations or with automation

that lacks structure (e.g., neural

networks)

40

HAT Agent

41

Thank you!

42

Three papers to appear in the proceedings of at the 8th International Conference on Applied Human Factors and

Ergonomics (AHFE 2017).

• Shively, R. J., Lachter, J., Brandt, S. L., Matessa, M., Battiste, V., & Johnson, W. W., Why Human-Autonomy

Teaming?

• Brandt, S.L., Lachter, J., Russell, R., & Shively, R. J., A Human-Autonomy Teaming Approach for a Flight-Following

Task.

• Lachter, J., Brandt, S. L., Sadler, G., & Shively, R. J., Beyond Point Design: General Pattern to Specific

Implementations.

Papers on ELP:

• Meuleau, N., Plaunt, C., Smith, D., Smith, T., An Emergency Landing Planner for Damaged Aircraft. Twenty-First

Conference on Innovative Applications of Artificial Intelligence (IAAI-09), pg 114-121.

• Meuleau, N., Plaunt, C., Smith, D., Smith, T., The Emergency Landing Planner Experiment. ICAPS-11 Scheduling

and Planning Applications Workshop (SPARK) pg 60-67.